The aim of this work is to establish consistent training sets using available low resolution ancillary data from the study area, which may be land cover maps with low spatial resolution. An initial training set is formed by a set of points randomly chosen from the land cover classes existing in the ancillary data. In a first phase some initial improvement of the training sets is done using the Normalized Difference Vegetation Index (NDVI). Three approaches are then tested to improve the initial training set using respectively the Maximum Likelihood Classifier, a classifier based on Dempster-Shafer theory and another based on the Mahalanobis distance. The degrees of certainty and uncertainty obtained with the soft classifier associated to each class is used to eliminate from the training set pixels that are not representative of the classes or identify new more reliable training pixels. A subsequent classification of the entire scene is made with the reference datasets obtained with the different approaches and the accuracy of the results tested building a confusion matrix and computing the User's, Producer's and Overall Accuracy. The proposed methodology is applied to a case study.